Progressive Pseudo-Label Optimization for Point-Supervised Change Detection

arXiv:2609.02171 · cs.CV · Submitted 2026-09-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Progressive Pseudo-Label Optimization for Point-Supervised Change Detection".

Jane: Here's the gist of the paper. Point-supervised change detection (PS-CD) aims to identify pixel-level changes between bi-temporal images using only sparsely annotated points Although point annotations substantially reduce labeling costs, their limited spatial coverage often results in incomplete and noisy pseudo-labels.

Tom: Framework Overview The proposed method introduces a two-stage framework that progressively adapts SAM2 priors to the target task In Stage I, SAM2 generates object-aware candidate masks from point annotations on the bi-temporal images, and a bi-temporal mask selection strategy is designed to convert generic segmentation responses into more reliable change pseudo-labels.

Jane: Subsequently, a lightweight CNN refinement module with an uncertainty-aware loss is employed to improve boundary quality and local structural consistency.

Tom: In Stage II, a teacher-student self-training framework is constructed where the teacher is updated by exponential moving average and periodically refreshes the pseudo-labels.

Paper summary: Tom: So, to recap what we just covered about "Progressive Pseudo-Label Optimization for Point-Supervised Change Detection," the core thesis is that point supervision gives us cheap labels, but those labels are noisy and incomplete, so this paper proposes a two-stage framework to fix that.

Jane: That two stages involve using SAM2 priors from a foundation model to generate candidate masks from those sparse points on both images, followed by a selection strategy to pick the most promising candidates.

Lu: And then Stage I uses those best candidates and refines them with a CNN-based module using an uncertainty-aware loss to suppress unreliable supervision and improve boundary quality.

Meng: While we're in Stage I, it’s important to note that this refinement isn't just about making things look cleaner; it’s about actively suppressing the errors from the initial sparse labels.

Lalam: Then Stage II kicks in with a teacher-student self-training process where the teacher model is updated via exponential moving average, and it periodically refreshes those pseudo-labels based on what the student model predicts.

Tom: So, it’s a progressive adaptation of SAM2 priors into change detection by first generating candidates and refining them, then iteratively improving those labels through self-training. It addresses the central questions about converting foundation model priors into reliable bi-temporal change pseudo-labels and how to make them better during training.

Jane: The paper claims this two-stage framework is effective because it systematically tackles the limitations of relying on local appearance similarity or coarse discriminative responses that plague other pseudo-label strategies like region expansion or CAM propagation, which often lead to incomplete regions and inaccurate boundaries.

Lu: They’re moving away from treating pseudo-labels as fixed supervision, allowing early errors to propagate through training, and instead they are iteratively improving them.

Meng: I need to be clear about the practical implication: this is a step toward making weakly supervised methods more robust for real-world deployment by ensuring the labels we use in Stage II aren't just static guesses from Stage I.

Lalam: The paper shows that the full model achieves top results across three benchmark datasets, specifically mentioning eighty-one point two five percent F1 and sixty-eight point eight zero percent IoU on LEVIR-CD for instance, which gives us some concrete performance numbers to look at when comparing it against other weak methods.

Tom: Those numbers are solid, especially since they state that this method achieves the best overall performance among weakly supervised approaches on most datasets tested. That’s a strong claim when you're dealing with limited supervision.

Jane: So, what does this mean for the world is that we can start leveraging powerful vision models for change detection even when our data isn't perfectly labeled yet, pushing the boundary of what’s possible in remote sensing applications.

Conclusion: Tom: Let's wrap up our discussion on "Progressive Pseudo-Label Optimization for Point-Supervised Change Detection." The authors are building this two-stage framework to take those initial, noisy point labels and progressively optimize them through SAM2 priors and iterative self-training.

Jane: The paper essentially shows that the power isn't just in the initial generation, but in the refinement process where Stage I cleans up the noise, and then Stage II keeps making those labels better through continuous learning.

Lu: The authors are essentially answering how to convert those general foundation model priors into reliable change pseudo-labels and how to improve them during training rather than leaving them fixed.

Meng: From a deployment view, this means we need to be careful about the complexity because you’re dealing with two stages of optimization that must work in harmony.

Lalam: The paper confirms that they found the best refresh interval for their self-training is every two epochs, which gives us a specific tuning knob for how often to update those pseudo-labels.

Tom: It boils down to using point supervision as a starting point and then applying structured optimization—the selection strategy, the uncertainty loss, and the teacher-student framework—to build better pseudo-labels over time.

Jane: The implication is that we can get much higher quality change detection results even with very little initial

Hailong Ning, Hao Wang, Yimeng Wang, Tao Lei, Renwei Dian, Asoke K. Nandi

Xi’an University of Posts and Telecommunications · Dalian Maritime University · School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology · School of Robotics, Hunan University · Department of Electronic and Electrical Engineering, Brunel University of London

cs.CV

Submitted: 2026-09-02

Updated: 2026-11-02

Comments: 9 pages,4 figures

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 82/100

The gist: Point-supervised change detection (PS-CD) aims to identify pixel-level changes between bi-temporal images using only sparsely annotated points Although point annotations substantially reduce labeling…

Key concepts

SAM2
SAM2 is a model used to generate candidate masks based on point prompts in bi-temporal images. It helps create multiple potential object masks for each timestamp, providing the initial spatial guidance needed for change detection tasks.
Bi-Temporal Mask Selection
This strategy ranks the candidate masks generated by SAM2 using a weighted score that considers SAM confidence, change consistency, and point coverage. This ensures that only the most reliable initial masks are selected to serve as high-quality pseudo-labels for change detection.
Uncertainty-Aware Loss
This loss function is used during the CNN refinement stage to suppress unreliable supervision. It specifically penalizes predictions where the model is uncertain, helping to filter out noisy or inaccurate initial pseudo-labels and improve boundary quality.
Teacher-Student Self-Training (EMA)
In Stage II, a teacher model updates its parameters using an exponential moving average (EMA) of the student's parameters. This updated teacher then generates new pseudo-labels to refresh the training data, creating a closed loop that progressively improves label quality and task adaptation.

Terminology

Summary

Point-supervised change detection (PS-CD) aims to identify pixel-level changes between bi-temporal images using only sparsely annotated points Although point annotations substantially reduce labeling costs, their limited spatial coverage often results in incomplete and noisy pseudo-labels. p. 0

Where the paper says it

uk Abstract Point-supervised change detection (PS-CD) aims to identify pixel-level changes between bi-temporal images using only sparsely annotated points. Although point annotations substantially reduce labeling costs, their limited spatial coverage often results in incomplete and noisy pseudo-labels.

Framework Overview

The proposed method introduces a two-stage framework that progressively adapts SAM2 priors to the target task In Stage I, SAM2 generates object-aware candidate masks from point annotations on the bi-temporal images, and a bi-temporal mask selection strategy is designed to convert generic segmentation responses into more reliable change pseudo-labels. Subsequently, a lightweight CNN refinement module with an uncertainty-aware loss is employed to improve boundary quality and local structural consistency. p. 0

Where the paper says it

In Stage I, SAM2 generates object-aware candidate masks from point annotations on the bi-temporal images, and a bi-temporal mask selection strategy is designed to convert generic segmentation responses into more reliable change pseudo-labels. Subsequently, a lightweight CNN refinement module with an uncertainty-aware loss is employed to improve boundary quality and local structural consistency.

In Stage II, a teacher-student self-training framework is constructed where the teacher is updated by exponential moving average and periodically refreshes the pseudo-labels. p. 0
Where the paper says it

Subsequently, a lightweight CNN refinement module with an uncertainty-aware loss is employed to improve boundary quality and local structural consistency. In Stage II, we construct a teacher-student self-training framework in which the teacher is updated by exponential moving average and periodically refreshes the pseudo-labels.

Stage I Pseudo-Label Generation and Refinement

Stage I focuses on generating initial soft pseudo-labels from sparse supervision This stage involves several critical steps to improve the quality of the initial predictions. The process includes:

  1. SAM2-based Candidate Mask Generation, where SAM2 is applied to each point prompt to generate multiple candidate masks for each timestamp. p. 2

    Where the paper says it

    For each prompt point pk, SAM2 is independently applied to the two temporal images to generate multiple candidate masks. The generated masks at the two timestamps are denoted as the candidate mask sets Mt1 k and Mt2 k , respectively, where each set contains N candidate masks corresponding to the same prompt point.

  2. Bi-Temporal Mask Selection, which ranks candidate masks using a weighted score based on SAM confidence, change consistency, and point coverage. p. 2

    Where the paper says it

    Step 2: Bi-Temporal Mask Selection. Since SAM2 may generate noisy or over-expanded masks, we rank candidate masks using a weighted score based on three criteria: SAM2 confidence s sam k,n , change consistency s chg k,n , and point coverage s pt k,n.

    The selection score is formulated as skk,n = λ1s sam k,n + λ2s chg k,n + λ3s pt k,n. p. 3
    Where the paper says it

    ferences between bi-temporal images: D = Norm(Dint + Dgrad + Dhf ), (1) s chg k,n = P x∈Ω Mk,n(x)D(x) P x∈Ω Mk,n(x) + ϵ , (2) s pt k,n = 1 |Pk| X p∈Pk I[Mk,n(p) > 0], (3) where Norm(·) denotes min-max normalization, Pk represents the point prompts of the k-th component, and I(·) is the indicator function. The final score and pseudo-label generation process are formulated as sk,n = λ1s sam k,n + λ2s chg k,n + λ3s pt k,n, Y˜ (x) = max k (skMk(x)), (4) where λ1, λ2, λ3 balance the three criteria.

  3. CNN-based Pseudo-Label Refinement, where a siamese encoder-decoder architecture is used with an uncertainty-aware loss to suppress unreliable supervision. p. 3

    Where the paper says it

    Therefore, we further train a CNN-based change detector using the generated soft pseudo-labels as supervision. Specifically, we adopt a siamese encoder-decoder architecture, where a shared ResNet-18 backbone extracts multi-scale features from the two temporal images, and the absolute feature differences are progressively decoded to produce a dense change probability map.

    The refinement loss is formulated as Lref = Lw bce(P, Y˜) + Lw dice(P, Y˜). p. 3
    Where the paper says it

    Inspired by uncertainty-aware pseudo-label learning, we assign both image-level and pixellevel confidence weights to the soft pseudo-labels, so that reliable regions contribute more to optimization while ambiguous regions are suppressed. w(x) = 1 − E(Y˜ ) 2Y˜ (x) − 1 2 , (8) E(Y˜ ) = 1 |Ω| X x∈Ω H Y˜ (x) , (9) Lref = L w bce(P, Y˜ ) + L w dice(P, Y˜ ), (10) where H(·) denotes the binary entropy function, L w bce and L w dice denote the weighted binary cross-entropy loss and weighted Dice loss under the uncertainty-aware weight w(x), respectively.

Stage II Iterative Optimization

Stage II establishes a closed-loop optimization process through a teacher-student self-training framework Instead of treating pseudo-labels as fixed supervision, the teacher model is updated using the exponential moving average (EMA) of the student parameters. The updated teacher produces more reliable predictions to refresh the pseudo-labels. p. 3

Where the paper says it

The updated teacher produces more reliable predictions to refine the pseudo-labels: Y˜ r+1 = µY˜ r + (1 − µ)Pt, (14) Lstage2 = Lu(Ps, Y˜ r+1). (15) where Y˜ r and Y˜ r+1 denote the pseudo-labels before and after the r-th refresh step, respectively, and µ is the momentum coefficient for pseudo-label updating.

The update rule for the teacher is θt ← αθt + (1 − α)θs, and the pseudo-label refresh is defined as Y˜ r+1 = µY˜ r + (1 − µ)Pt. p. 3
Where the paper says it

We introduce a teacher-student self-training framework. Given the pseudolabels generated in Stage I, the student network is first optimized to learn bi-temporal change representations: Ps = Ψs(I t1 , It2 ; θs), (11) Pt = Ψt I t1 , It2 ; θt , (12) θt ← αθt + (1 − α)θs, (13) where Ψs(·) denotes the student network, Ψt(·) denotes the teacher network, and θs and θt are their corresponding parameters.

Experimental Validation and Results

Experiments on three benchmark datasets—WHU-CD, LEVIR-CD, and SYSU-CD—demonstrate the effectiveness of the proposed method. Quantitative analysis shows that the method achieves the best overall performance among weakly supervised approaches on most datasets. p. 4

Where the paper says it

Quantitative analysis. From table 1, it can be observed that our method achieves the best overall performance among weakly supervised approaches on most datasets, demonstrating the effectiveness of introducing SAM priors and the proposed two-stage closed-loop optimization framework.

For instance, on LEVIR-CD, our method achieves 81. p. 4
Where the paper says it

It is worth noting that, although our method is trained under weak supervision, its performance is already close to, or even better than, some fully supervised methods on several datasets. 1) Results on the LEVIR-CD: our method achieves 81.

25% F1 and 68.80% IoU. Qualitative analysis confirms that the proposed approach produces more complete change regions and clearer boundaries compared to existing methods. p. 4
Where the paper says it

Figures 3 show qualitative comparisons with both fully supervised and weakly supervised methods on benchmark datasets. Compared with existing methods, the proposed approach produces more complete change regions, clearer boundaries, and fewer false alarms.

The full model achieves the best results on all three datasets, demonstrating that Stage I refinement and Stage II self-training are complementary. p. 5
Where the paper says it

Adding self-training in Stage II also brings clear gains over the initial pseudo-label baseline, validating the benefit of iterative pseudo-label optimization. The full model achieves the best results on all three datasets, demonstrating that Stage I refinement and Stage II self-training are complementary.

Ablation Study Insights

Ablation studies confirm the contribution of key components to the framework's success. Removing the bi-temporal mask selection strategy leads to consistent performance drops, validating its effectiveness in generating more reliable initial pseudo-labels. p. 5

Where the paper says it

A pivotal part of Stage I is the bi-temporal mask selection strategy, which ranks and fuses SAM-generated candidate masks by jointly considering SAM confidence, change consistency, and point coverage. As shown in Table 3, removing this strategy leads to consistent performance drops, validating its effectiveness in generating more reliable initial pseudo-labels.

Similarly, replacing the uncertainty-aware loss with a standard loss results in a significant performance decline, indicating that uncertainty-aware supervision better suppresses noisy pseudo-labels. p. 5
Where the paper says it

We further evaluate the role of the uncertainty-aware weighting scheme by replacing it with a standard loss without confidence weighting. The performance decline indicates that uncertainty-aware supervision can better suppress noisy pseudo-labels and guide the model to focus on more reli-

The EMA update strategy in Stage II is also shown to be important for progressively improving pseudo-label quality and task-specific adaptation. p. 6
Where the paper says it

To validate the role of the EMA update strategy, we replace the EMA-updated teacher with a non-EMA variant while keeping the rest of the self-training framework unchanged. The inferior results indicate that EMA updating provides more stable supervision and is important for progressively improving pseudo-label quality and task-specific adaptation.

The best performance is achieved when the pseudolabels are refreshed every 2 epochs, suggesting a moderate refresh interval provides the best balance between stability and adaptivity. p. 6
Where the paper says it

In contrast, overly sparse refreshing also hurts performance because outdated pseudo-labels limit the benefit of self-training. These results suggest that a moderate refresh interval provides the best balance between stability and adaptivity.

Conclusion

In conclusion, the proposed two-stage point-supervised change detection framework exploits foundation-model priors and iterative pseudo-label optimization. SAM2 is used to generate candidate masks from bi-temporal point supervision, and a bi-temporal mask selection strategy together with CNN-based refinement is introduced to construct higher-quality soft pseudo-labels. p. 6

Where the paper says it

01 Table 5: Sensitivity analysis of the pseudo-label refresh interval on the LEVIR-CD dataset. SAM2 is used to generate candidate masks from bi-temporal point supervision, and a bi-temporal mask selection strategy together with CNN-based refinement is introduced to construct higher-quality soft pseudo-labels.

In Stage II, a teacher-student self-training framework is further developed to progressively refresh pseudo-labels and improve taskspecific adaptation. p. 6
Where the paper says it

SAM2 is used to generate candidate masks from bi-temporal point supervision, and a bi-temporal mask selection strategy together with CNN-based refinement is introduced to construct higher-quality soft pseudo-labels. In Stage II, a teacher-student self-training framework is further developed to progressively refresh pseudo-labels and improve taskspecific adaptation.

Extensive experiments on three benchmark datasets show that the proposed method achieves the best results among the compared weakly supervised approaches, while remaining competitive with several fully supervised baselines. p. 6
Where the paper says it

In Stage II, a teacher-student self-training framework is further developed to progressively refresh pseudo-labels and improve taskspecific adaptation. Extensive experiments on three benchmark datasets show that the proposed method achieves the best results among the compared weakly supervised approaches, while remaining competitive with several fully supervised baselines.

In future work, we will further study how to improve pseudo-label reliability on challenging scenes and extend the proposed framework to more general forms of weak supervision. p. 6
Where the paper says it

Extensive experiments on three benchmark datasets show that the proposed method achieves the best results among the compared weakly supervised approaches, while remaining competitive with several fully supervised baselines. In future work, we will further study how to improve pseudo-label reliability on challenging scenes and extend the proposed framework to more general forms of weak supervision.

References

Bandara, W. G. C., 2022, A Transformer-Based Siamese Network for Change Detection

Caye Daudt, R.; Le Saux, B. p. 5

Where the paper says it

Method LEVIR-CD WHU-CD SYSU-CD P R F1 IoU OA P R F1 IoU OA P R F1 IoU OA Fully-supervised Methods FC-Siam-Conc (Caye Daudt, Le Saux, and Boulch 2018) 90. 49 FC-Siam-Diff (Caye Daudt, Le Saux, and Boulch 2018) 91.

; and Boulch, A. p. 5
Where the paper says it

Method LEVIR-CD WHU-CD SYSU-CD P R F1 IoU OA P R F1 IoU OA P R F1 IoU OA Fully-supervised Methods FC-Siam-Conc (Caye Daudt, Le Saux, and Boulch 2018) 90. 49 FC-Siam-Diff (Caye Daudt, Le Saux, and Boulch 2018) 91.

, 2018, Fully Convolutional Siamese Networks for Change Detection

Ding, L.; Zhu, K.; Peng, D.; Tang, H.; Yang, K.; and Bruzzone, L. p. 7

Where the paper says it

In 2018 25th IEEE International Conference on Image Processing (ICIP), 4063–4067. ; and Bruzzone, L.

, 2024. Adapting Segment Anything Model for Change Detection in VHR Remote Sensing Images

Dong, S.; Wang, L.; Du, B., and Meng, X. p. 7

Where the paper says it

IEEE Transactions on Geoscience and Remote Sensing, 62: 1–11. ; and Meng, X.

, 2024. ChangeCLIP: Remote sensing change detection with multimodal vision-language representation learning

Fang, L.; Jiang, Y. p. 1

Where the paper says it

To reduce annotation burden, recent studies (Fang et al. 2023; Jiang et al.

; Yu, H.; Zhang, Y. p. 1
Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

; and Yue, J. p. 7
Where the paper says it

; Zhang, Y. ; and Yue, J.

, 2023. Point Label Meets Remote Sensing Change Detection: A Consistency-Aligned Regional Growth Network

Fang, S.; Li, K.; Shao, J., and Li, Z., 2021. SNUNet-CD: A Densely Connected Siamese Network for Change Detection of VHR Images

Han, C., 2023. HCGMNET: A Hierarchical Change Guiding Map Network For Change Detection

Han, C.; Wu, C.; Guo, H.; Hu, M.; Li, J.; and Chen, H. p. 1

Where the paper says it

Remote sensing change detection has achieved remarkable progress with the development of deep neural networks. However, most existing methods (Hao Chen and Shi 2021) rely on dense pixel-level annotations, which are costly and labor-intensive to obtain for large-scale remote sensing imagery.

, 2023. Change Guiding Network: Incorporating Change Prior to Guide Change Detection in Remote Sensing Imagery

Hao Chen, Z. Q., and Shi, Z. p. 1

Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

, 2021. Remote Sensing Image Change Detection with Transformers

He, C.; Zhang, R. p. 1

Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

; Tang, L.; Yang, Z.; Li, K.; Fan, D.-P.; and Farsiu, S. p. 7
Where the paper says it

; Zhang, R. ; and Farsiu, S.

, 2025. SCALER: SAM-Enhanced Collaborative Learning for Label-Deficient Concealed Object Segmentation

Jiang, F., Zhong, Z. p. 7

Where the paper says it

arXiv:2511. ; Zhong, Z.

; Zhang, M. p. 1
Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

p. 1
Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

; Gong, M.; Zhou, Y.; Zhao, W.; and Guan, Z. p. 7
Where the paper says it

; Zhang, M. ; and Guan, Z.

, 2025. BARNet: Boundary-Aware Refinement Network for Weakly Supervised Change Detection

Kage, P., Rothenberger, J. p. 7

Where the paper says it

IEEE Transactions on Geoscience and Remote Sensing. ; Rothenberger, J.

, Andreadis, P. p. 7
Where the paper says it

; Rothenberger, J. ; Andreadis, P.

, and Diochnos, D. p. 7
Where the paper says it

; Andreadis, P. ; and Diochnos, D.

, 2026. A Review of Pseudo-Labeling for Computer Vision

Kirillov, A.; Mintun, E. p. 7

Where the paper says it

Kirillov, A. ; Mintun, E.

; Ravi, N.; Mao, H.; Rolland, C. p. 7
Where the paper says it

; Mintun, E. ; Rolland, C.

; Gustafson, L. p. 7
Where the paper says it

; Rolland, C. ; Gustafson, L.

; Xiao, T., et al., 2023. Segment Anything

Lei, T., Xu, Y.; Ning, H., Lv, Z., Min, C., Jin, Y., and Nandi A. p. 0

Where the paper says it

cn, drw@hnu. nandi@brunel.

p. 0
Where the paper says it

cn, drw@hnu. nandi@brunel.

K., 2024. Lightweight Structure-Aware Transformer Network for Remote Sensing Image Change Detection

Li, Z.; Tang, C.; Liu, X.; Li, C.; Li, X.; and Zhang W. p. 1

Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

, 2023. MS-Former: Memory-Supported Transformer for Weakly Supervised Change Detection with Patch-Level Annotations

Liu, N., Xu, X.; Su, Y.; Zhang H., and Li H.-C., 2025a. PointSAM: Pointly-Supervised Segment Anything Model for Remote Sensing Images

Liu, T.; Xu, J.; Lei T.; Wang Y., Du X., Zhang W., Lv Z., and Gong M. p. 7

Where the paper says it

; Zhang, W. ; and Gong, M.

, 2025b. AEKAN: Exploring SuperpixelBased AutoEncoder Kolmogorov-Arnold Network for Unsupervised Multimodal Change Detection

Meyer, A.; Mutter, D. p. 1

Where the paper says it

Similar strategies have also been successfully applied to camouflaged object detection (He et al. 2025) and medical object detection (Meyer, Mutter, and Padoy 2026), where progressively improved pseudo-labels provide effective supervision under limited annotations.

, and Padoy, N. p. 1
Where the paper says it

Similar strategies have also been successfully applied to camouflaged object detection (He et al. 2025) and medical object detection (Meyer, Mutter, and Padoy 2026), where progressively improved pseudo-labels provide effective supervision under limited annotations.

, 2026. DExTeR: Weakly Semi-Supervised Object Detection with Class and Instance Experts for Medical Imaging

Ning, H.; He, Q.; Lei T.; Cao X.; Zhang W.; Chen Y., and Nandi A. K., 2025. DA2-Net: Integrating SAM2 With Domain Adaption and Difference Aggregation for Remote Sensing Change Detection

Subhani, M. N., 2026. ReSAM: Refine, Requery, and Reinforce: Self-Prompting Point-Supervised Segmentation for Remote Sensing Images

Wang, L.; Zhang, M., and Shi W. p. 1

Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

, 2023. CS-WSCDNet: Class Activation Mapping and Segment Anything ModelBased Framework for Weakly Supervised Change Detection

Wang, Y.; Li E.; Samat A.; Liu W., and Li X., 2026. PGU-CD: A Point-Guided Uncertainty-Aware Framework for Building Change Detection

Wei, C., 2025. ASS-CD: Adapting Segment Anything Model and Swin-Transformer for Change Detection in Remote Sensing Images

Wu, C., 2022. Fully Convolutional Change Detection Framework with Generative Adversarial Network for Unsupervised, Weakly Supervised and Regional Supervised Change Detection

Zhao, Z.; Ru L.; Wu C., and Wang D. p. 1

Where the paper says it

ate pseudo-labels through region expansion (Fang et al. 2023) or class activation maps (CAMs) propagation (Wang, Zhang, and Shi 2023).

, 2025.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Make SAM2-guided pseudo-label generation more robust by incorporating an entropy-based image level weighing mechanism to address noisy or over-expanded masks generated by SAM2, ensuring that reliable regions contribute more to optimization while ambiguous regions are suppressed.

  2. Develop a bi-temporal mask selection strategy that jointly considers SAM confidence, change consistency, and point coverage by using the defined score formula (Equation 4) with carefully tuned weights lambda1, lambda2, and lambda3 to transform generic single-temporal segmentation responses into more reliable change pseudo-labels.

  3. Enhance boundary quality by implementing a CNN-based Pseudo-Label Refinement module that uses an uncertainty aware loss (Equation 10) weighted by the uncertainty weight w(x), ensuring that reliable regions contribute more to optimization while ambiguous regions are suppressed in the refinement stage.

  4. Establish a closed-loop learning process where the teacher model is updated using exponential moving average (EMA) of student parameters and periodically refreshes pseudo-labels, enabling iterative pseudo-label optimization by alternating between pseudo-label refinement and model reoptimization.

  5. Improve task specific adaptation by using the EMA update strategy for the teacher network in Stage II, as this design enables the teacher to accumulate the task-adaptive knowledge learned by the student and to provide more stable guidance during training.

  6. Increase pseudo-label reliability through periodic pseudo-label refreshing in Stage II, where the teacher produces more reliable predictions to refine the pseudo-labels, allowing for continuous improvement of supervision during training.

Abstract

Point-supervised change detection (PS-CD) aims to identify pixel-level changes between bi-temporal images using only sparsely annotated points. Although point annotations substantially reduce labeling costs, their limited spatial coverage often results in incomplete and noisy pseudo-labels. To address this issue, we propose a two-stage framework that introduces SAM2 priors into PS-CD and progressively adapts them to the target task. In Stage I, SAM2 generates object-aware candidate masks from point annotations on the bi-temporal images, and a bi-temporal mask selection strategy is designed to convert generic segmentation responses into more reliable change pseudo-labels. Subsequently, a lightweight CNN refinement module with an uncertainty-aware loss is employed to improve boundary quality and local structural consistency. In Stage II, we construct a teacher-student self-training framework in which the teacher is updated by exponential moving average and periodically refreshes the pseudo-labels. This design establishes a closed-loop optimization process that alternates between pseudo-label refinement and model re-optimization. Experiments on three benchmark datasets, including WHU-CD, LEVIR-CD, and SYSU-CD, demonstrate that the proposed method outperforms previous weakly supervised approaches on most benchmarks and remains competitive with several fully supervised methods.

Sources

Related papers